Pathological features can be learned through hierarchical representation modeling for Alzheimer’s disease staging

Abstract Accurate staging of Alzheimer’s disease (AD) is a pressing challenge due to the continuum of pathological features that range from non-demented to moderately demented states. In this study, we present a hierarchical representation modeling framework that uses a depthwise separable squeeze and excitation network (DWSENet) to classify four clinically meaningful categories from MRI: Non-Demented, Very Mild Demented, Mild Demented, and Moderately Demented. A total of 44,000 images were curated from the Kaggle Alzheimer’s dataset via augmentation. The balanced training set contained 32,000 images (8000 per class) and an independent test set included 8000 images (2000 per class). DWSENet achieved per-class precision of 0.997 (Non-Demented), 0.978 (Very Mild), 0.990 (Mild), and 0.982 (Moderate). Most importantly, no Moderately Demented cases were misclassified. The Matthews Correlation Coefficient reached 0.982 with a narrow 95% confidence interval ranging from [0.979, 0.985]. Comparative evaluation against logistic regression, LeNet, and a multilayer perceptron (MLP) demonstrated statistically significant improvements: mean accuracy differences of -0.3226 versus logistic regression ( p < 0.001) and − 0.0875 versus MLP ( p < 0.001), though not significant versus LeNet (-0.0224, p = 0.1314). Receiver operating characteristic analysis yielded class-specific AUC values approaching 1.00, while calibration curves confirmed close alignment between predicted and observed probabilities. These results highlight DWSENet’s clinical relevance as a tool for automated AD staging and early intervention planning.